Deep Learning-Based Modulation Recognition with Multi-Scale Temporal Feature Extraction

This paper studies a deep learning-driven method for identifying modulation types in communication signals without prior information. To effectively capture the relationship between points in the time-series data, we apply a multi-head self-attention mechanism. The results demonstrate superior modulation recognition accuracy compared to conventional AMR models. In particular, the classification accuracy between QAM16 and QAM64 was significantly improved, confirming that the performance of the model has been enhanced compared to previous models.